[CI] Bound the CUDA graph capture range in test launches and lift the spec fixture's admission cap (#33776)

test/registered/cpu/test_spec_eagle_cpu.py

@@ -40,6 +40,8 @@ class _Core(EagleLlama2Base):
     attention_backend = "intel_amx"
     disable_overlap = True
     mem_fraction_static = 0.3
+    # CPU decode is compute-bound; a wider batch buys nothing here.
+    max_running_requests = 8
     gsm8k_num_examples = 64
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
 

test/registered/cpu/test_spec_eagle_parity_cpu.py

@@ -20,6 +20,8 @@ class TestEagle3ParityCPU(SpecParityKit, Eagle3Base):
     attention_backend = "intel_amx"
     disable_overlap = True
     mem_fraction_static = 0.3
+    # CPU decode is compute-bound; a wider batch buys nothing here.
+    max_running_requests = 8
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
 
 

test/registered/cpu/test_spec_eagle_topk_cpu.py

@@ -32,6 +32,8 @@ class _Core(EagleLlama2Base):
     attention_backend = "intel_amx"
     disable_overlap = True
     mem_fraction_static = 0.3
+    # CPU decode is compute-bound; a wider batch buys nothing here.
+    max_running_requests = 8
     gsm8k_num_examples = 64
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)